Subsequent Leap to Harness Engineering: JiuwenClaw Pioneers ‘Coordination Engineering’

The right way to make a number of brokers work collectively like an elite crew — autonomously dividing duties, speaking effectively, and collaborating seamlessly?

The openJiuwen neighborhood launched the newest model of JiuwenClaw, which provides help for AgentTeam — a multi-agent collaborative functionality. It proposes that the following leap past Harness Engineering is Coordination Engineering.

In in-depth checks, this crew collaboration mechanism has demonstrated exceptional stability —crew members have clear roles, collaborate autonomously with seamless coordination, and your entire workflow requires no human intervention.

How hardcore is it, actually?

It could possibly autonomously assemble a “well-trained” crew of brokers — and with that crew, it could actually produce a strong, logically rigorous 200‑web page technical PPT in underneath 20 minutes.

Challenge hyperlinks: https://github.com/openJiuwen-ai/jiuwenclaw

Testing JiuwenClaw “Coordination Engineering” in Motion

Need deep insights with out lifting a finger? A 200‑web page, content material‑wealthy PPT in underneath 20 minutes.

In our trial, we requested it to conduct an in‑depth investigation of OpenClaw expertise and break it down throughout 10 core elements. For every side, it assigned a devoted agent to take cost. Every agent was accountable for producing 20 PPT slides, all underneath a unified theme. Lastly, the ten units of slides have been merged into a whole, 200‑web page technical presentation.

Your entire course of took lower than 20 minutes. The ensuing PPT was detailed, logically structured, and impressively environment friendly.

Technical Breakdown: Three Core Capabilities of JiuwenClaw AgentTeam

The core design philosophy of AgentTeam is easy: simulate how real-world groups collaborate.

• A Chief Agent is accountable for requirement evaluation, crew constructing, and process planning.
• A number of Teammate Brokers declare duties, execute independently, report outcomes, and collaborate by means of a shared workspace.
• Throughout execution, key milestones require Chief approval, and fault restoration is automated.

 1. Hierarchical Autonomous Collaboration: Chief Orchestrates Intelligently, Teammates Execute Autonomously

JiuwenClaw AgentTeam delegates this duty to the Chief Agent itself.

What the Chief does:

• Dynamically builds the crew: Assigns roles and members dynamically primarily based on the purpose. If extra arms are wanted mid-execution, it could actually add or take away members on the fly.
• Plans duties: Breaks down the purpose into concrete duties, establishing dependencies (e.g., “evaluation can solely begin after knowledge assortment is full”).
• Assigns and screens: After creating duties, it tracks progress in actual time—who claimed what, who accomplished what, who bumped into points—and adjusts accordingly.

What Teammates do:

• Declare duties proactively: Browse the duty board and declare duties that match their capabilities.
• Execute independently: Full their work inside their very own workspace. 
• Report outcomes: Replace the standing and notify the Chief and different dependents.

Crew members drive the core workflow by means of process collaboration—claiming, executing, finishing, unblocking downstream duties—discussing plans, negotiating priorities, flagging points, requesting help.

Each channels run in parallel, with process dependencies managed routinely—not merely mechanical distribution and aggregation.

2. Crew Workspace: A Shared Crew File House

JiuwenClaw AgentTeam solves this with Crew Workspace—a crew‑stage shared file area that every one members can transparently entry. Every Teammate’s working listing routinely mounts a shared path pointing to the identical crew workspace.

3. Full Lifecycle Administration: From Plan Approval to Automated Fault Restoration

3.1 Chief Approval

AgentTeam gives a two‑layer approval mechanism:

• Plan mode: For essential duties, a Teammate first submits an execution plan for Chief approval. 
• Instrument approval: When a Teammate must carry out a delicate operation (e.g., deleting information, calling exterior APIs, modifying shared configurations), Chief approval is required.

3.2 Occasion‑Pushed Mechanism

AgentTeam mitigates this with an occasion‑pushed mechanism, utilizing each exterior and inside occasions:

• Exterior occasions: Job state adjustments, member lifecycle adjustments, inter‑member messages—any significant change triggers an occasion.
• Inner occasions: Framework‑generated self‑test occasions (mailbox polling, process board polling) act as a security internet.

After an occasion is triggered, the related brokers are routinely woke up (e.g., idle Teammates declare duties, the Chief reassigns timed-out duties)

3.3 Persistent Groups

With Persistent mode enabled, groups will be preserved throughout periods: The subsequent time you want the crew, you possibly can restore it with one click on—create a brand new session area, restart the crew members, and also you’re able to go, with out rebuilding the crew from scratch.

3.4 TeamMonitor

TeamMonitor offering observability in two dimensions:

• Question API: Verify crew info, member states, process progress, and different statuses at any time.
• Occasion stream: Subscribe to crew occasions in actual time. Job completions, member state adjustments, messages despatched/acquired… all occasions will be consumed one after the other through an asynchronous iterator. You’ll be able to construct dashboards, logging techniques, or set off exterior workflows from these occasions. Each step of the crew’s operation is traceable and auditable.

Core Underpinning: openJiuwen AgentTeam Structure

The core technical rules of AgentTeam will be summarized in three factors:

1. Constant collaboration through a shared process listing: All members share the identical dynamic process listing. Every agent independently claims and executes duties primarily based on the crew purpose, process definitions, and its personal capabilities—guaranteeing pure info consistency.
2. Twin‑drive mannequin of messages and duties: Members drive the core workflow by means of process transitions, whereas additionally repeatedly discussing and negotiating through a message channel exterior the duty system—masking every part from structured execution to unstructured communication.
3. Function and gear engineering: RolePolicy defines the behavioral norms and choice boundaries of the Chief and Teammates throughout the crew. TeamTools endows crew members with particular coordination capabilities. The position determines “what must be carried out,” and the instruments decide “what will be carried out.”

About JiuwenClaw

JiuwenClaw is a “Claw” Agent developed on high of the openJiuwen open‑supply neighborhood. It natively helps multi‑agent collaboration and agent self‑evolution. The core design philosophy is easy: Perceive what you need, and evolve autonomously.

Past AgentTeam, JiuwenClaw can be very straightforward to put in and deploy – a single command will get you up and working. For a fast begin, consult with: https://github.com/openJiuwen-ai/jiuwenclaw/blob/develop/docs/en/Quickstart.md

As well as, JiuwenClaw presents a number of benefits in autonomous process planning, self‑evolution, context compression and offloading, browser manipulation, and general “lobster‑like” dealing with:

• Autonomous process administration: at all times prepared if you end up : JiuwenClaw includes a process planning mode, which is basically a to‑do listing for the AI. Customers can dynamically interrupt, append, or modify duties at any time. 
• Self‑evolving Expertise: Proactively data these execution errors and suggestions, analyzes the basis trigger, and generates focused enchancment solutions. An evolution approval window then pops up for the person – each replace is your name.
• Context compression & offloading : Successfully reduces prices by managing context size.
• Layered Reminiscence: Achieves long-term storage and clever retrieval of situations and operation traces. 
• Browser manipulation: Robotically accesses profile info like cookies and native cache, seamlessly taking up the browser setting. 

About OfficeClaw

The enterprise-grade model, OfficeClaw, constructed on the Harness engineering basis, seamlessly integrates process planning, multi-agent collaboration, instrument invocation, and safety governance on Huawei Cloud AgentArts, bettering the success price of advanced workplace duties.

Be part of the Neighborhood & Discover openJiuwen

openJiuwen Obtain Hyperlinks

JiuwenClaw Obtain Hyperlinks

• JiuwenClaw on GitHub: https://github.com/openJiuwen-ai/jiuwenclaw
• JiuwenClaw on AtomGit: https://gitcode.com/openJiuwen/jiuwenclaw
• AgentArts on Huawei Cloud：https://www.huaweicloud.com/product/agentarts
• OfficeClaw on Huawei Cloud：https://www.huaweicloud.com/product/agentarts/officeclaw.html

_{Observe: Because of the OpenJiuwen crew for the sources, pictures, video, and different particulars.}